From Frames to Worlds

A one-megapixel image is a million numbers; a minute of video is billions, and every one of them must agree with its neighbours in space and in time. That constraint — temporal coherence — is why video generation lagged image generation by years, and why solving it points somewhere bigger than clips: models that simulate worlds.

This book traces that arc chronologically, from blurry GAN loops to Sora-class diffusion transformers and action-conditioned world models, with the mathematics explained in plain English and every claim tied to its original paper.